The scariest agent failure isn't the one that crashes — it's the one that quietly does the wrong thing while looking fine. The finale of the production AI agents series: observability, ops, cost control, and the feedback loop that closes the whole system.
One agent can only do so much. Part 4 of the production AI agents series: when to reach for a team, the orchestration patterns that actually work, handoff protocols, and why the hard part of multi-agent systems is org design, not code.
The moment you give an AI agent shell access, credentials, or write permissions, everything changes. Part 3 of the production agents series: designing tools that fail safe, treating credentials as non-negotiable, and the confidence signal that tells you when to verify.
Most 'AI memory' is just retrieval, and retrieval isn't continuity. Part 2 of the production AI agents series: the two-layer memory system that turns a goldfish into a colleague, and the context-window economics nobody warns you about.
Every AI agent demo works. Then you ship it and it falls apart. This is the first in a 5-part series on the unglamorous 80% nobody puts in the demo — starting with architecture: deciding what you don't let the model do.
AI assistants are confidently wrong all the time. This portable custom instruction forces them to self-score every response — so you know when to ship and when to verify.
What a multi-hop PXE failure taught me about managing AI as a troubleshooting partner — not a magic solution.
I was checking switch cabling and thought — why am I doing this manually? I asked GitHub Copilot to do it. That question led down a rabbit hole of credential backends, hanging regex loops, and an AI that admitted it was being dumb.
I handed GitHub Copilot a /26 subnet and told it to plan IP addressing for a two-switch iBGP fabric with spine uplinks. Then we deployed it. Here's what happened.
What happens when you point an AI at two factory-fresh SONiC switches and say 'figure it out' — a real-world experiment in AI-assisted network discovery.